Don't just focus on Figure 02. The latest achievements of domestic robots have been revealed, showing off their strength!

In Beijing Yizhuang, there is actually a "gym" for humanoid robots.

The robots here are the first to achieve training on a treadmill, the kind that can run 6km in an hour:

This is just a warm-up, after running, they still have to climb stairs:

When going outdoors, a dedicated path for robots is also arranged, and they can't stop even in 30-degree high temperatures:

Occasionally, they also have to deal with the difficulty of complex terrains such as lawns and slopes:It is the pure electric full-size humanoid robot named "Tiangong," which was first launched in China in April this year, developed by the Beijing Embodied Intelligence Robotics Innovation Center.

When it first came out, Tiangong mainly focused on anthropomorphic running. In just a few months, its capabilities have grown significantly.

In addition to the running on the treadmill and walking on complex outdoor terrains shown above, Tiangong now has a large model, can converse in both Chinese and English, and can follow human commands to pick up objects.

Recently, the 2024 World Robot Conference is about to be held. Quantum Position has learned in advance that the evolved version of "Tiangong," integrating "not only" these abilities, will be unveiled at the conference.

Advertisement

Moreover, Tiangong will change its "appearance," be equipped with hands, and upgrade its large joints.

Regarding the capabilities of the evolved Tiangong, we had a chat with its motion control leader, Guo Yijie.

Tiangong has 42 degrees of freedom in total, with the support of a large model.

A summary of the upgraded Tiangong's functional parameters is as follows:

Compared to the previous version, as shown in the table below, the height is 163cm, and the weight has increased from the original 43kg to 56kg.The total degrees of freedom have increased to 42, compared to the previous single arm which had 3 degrees of freedom and has now increased to 7, and the neck has also added 3 degrees of freedom.

First-generation Tian Gong functional parameters

Previously, Tian Gong "had no hands," but now it is equipped with them. A single hand weighs 600g, has 6 degrees of freedom, tactile sensor accuracy within 0.3N, and the gripping force of a single finger is greater than 1kg.

Moreover, Tian Gong is also equipped with four all-scenario binocular structured light 3D cameras, high-precision six-dimensional force sensors, which can achieve 360-degree environmental perception.

By perceiving the environment, Tian Gong can adapt to complex terrains, move on grasslands, sandy areas, hills, and gravel, and cope with a terrain difference of 30cm is not a problem.

The stability when running has also been improved, with a speed of up to 6km/h.

In addition, another major upgrade of Tian Gong is the carrying of a large model, which has multimodal operation capabilities.Now it can speak and interact with humans, supporting both Chinese and English.

It can also follow human commands to pick up objects:

[Unfortunately, a video cannot be inserted here... You can check it on the Quantum Bit public account~]

And these are just a part of it, there will be more demonstrations and surprises at the World Robot Conference.

When Tian Gong made its debut, it mainly focused on the single ability of anthropomorphic running. From the current evolutionary version, combined with perceptual interaction, it can be said that Tian Gong has preliminarily formed an embodied intelligent body form.

Researching embodied intelligent planning, decision-making, and task execution is also the current focus of the team behind Tian Gong.

Creating an Embodied Intelligent Body

Based on embodied intelligence, in terms of technical implementation, the research team has focused on enhancing Tian Gong's visual perception capabilities.

Previously, Tian Gong walked in a "blind" state, needing to probe the ground with its feet, and now based on visual perception, facing a large terrain difference, Tian Gong can make actions such as raising its legs in advance to cope with it.In terms of specific methods, the team, based on reinforcement learning, has independently developed a motion skill learning method—Predictive Reinforcement Imitation Learning based on State Memory.

As previously introduced by Quantum Bit, this method integrates the advantages of traditional methods with high stability, as well as the strong generalization and environmental independence of reinforcement learning.

It solves the problem of poor positioning accuracy brought by reinforcement learning, and also solves the problem of poor adaptability to unstructured environments in model predictive control methods.

Guo Yijie, the person in charge of Tiangong Motion Control, also revealed to Quantum Bit that during the previous training, the team found that some networks may be easily disturbed by sensor drift during actual operation, sometimes showing unstable postures, and thus proposed this method.

After the first generation of Tiangong was released, the team added more historical state memory to Tiangong in the training in the following months, enabling it to estimate its current state and environmental terrain, thereby achieving better generalization effects.

Guo Yijie also stated that the current embodied intelligence needs to solve the problem of "Action" task planning and execution.

In terms of both the type and complexity of tasks, it can cover most of the tasks in human daily work and life. The variety of tasks it can do increases, and the brain can cooperate to achieve more complex and longer-term tasks.

For the implementation of complex task planning, he shared several technical routes:The article translates to:

Like Tesla's robots, the main approach is to collect data and then use the data for supervised learning, training the robot to automatically execute in a relatively single and fixed scenario. Although this method is relatively quick in terms of results, its generalization ability is somewhat poor.

Another method is to train through reinforcement learning in a simulation environment, which mainly relies on continuous trial and error and self-learning within an environment. The main problem encountered with this method is how to transfer from the simulation environment to the real physical scene. From the perceptual level to the specific physical interaction, there are significant differences between the simulation and the real scene.

Another method is to use large models to output some task points and then use traditional motion planning to execute these tasks.

The solution of the Tiangong R&D team is to integrate different methods, and the next step is to create a meta-skill library:

At this stage, I think it is necessary to expand the robot's skill library... including these methods, each of which can be used to solve different task scenarios. Therefore, each skill in the skill library may be implemented in different ways.

Solving the "commonality" problem of domestic robots

Let's talk about the company behind Tiangong, which Quantum Bit has introduced before.

Beijing Embodied Intelligent Robotics Innovation Center (hereinafter referred to as the Innovation Center), formerly known as Beijing Humanoid Robot Innovation Center, was established in November of last year and was jointly established by Xiaomi Robots, Ubtech, Jingcheng Mechanical and Electrical, Yizhuang Robots, and others.

They aim to solve the key common problems of embodied intelligent robots and avoid the process of domestic robots industry repeating the simple reinventing of the wheel.The Innovation Center has gathered a group of top scientists and engineers and has also taken the lead in establishing the Innovation Center Expert Committee and the Beijing Humanoid Robot Industry Alliance.

The Innovation Center Expert Committee is chaired by Academician Qiao Hong of the Chinese Academy of Sciences, with Zhu Songchun, the Dean of the General Research Institute, Huang Tiejun, the Chairman of the Zhiyuan Research Institute, and Wei Ran, the Chief Engineer of the Information and Communication Institute, serving as deputy directors.

In April of this year, the Innovation Center launched the "Tiangong" general robot mother platform, which is an open-source platform for embodied intelligent hardware.

The "Tiangong" platform can achieve flexible expansion of software and hardware functional modules, providing a series of open interfaces, and research institutions and robot-related enterprises can carry out secondary development based on the hardware and software functions of the "Tiangong" mother platform.

At that time, they announced that they would focus on relying on large model-driven exploration of general embodied intelligent platforms. Now, the multifunctional embodied intelligent body mother platform "Kaiwu" is gradually emerging and is being developed intensively.

The "Kaiwu" platform focuses on the large model and framework of embodied intelligence, focusing on the construction of key methods for multimodality, embodied intelligent simulation applications, and the construction of a complete set of tool chains.

Around "Kaiwu", the team is building a large-scale embodied intelligent dataset to support the training and tuning of visual language multimodal large models with more than 7 billion parameters, achieving capabilities such as Chinese interaction, open Q&A, scene visual understanding, and embodied operation for robots.

The plan is to release 2 million high-quality data by the end of 2025.

By the way, there have been many new developments in the field of humanoid robots recently.UBTECH has revealed that humanoid robots are already working in the Zeekr factory; Tesla's official images show that Optimus has entered the factory to pick up batteries; the OpenAI-empowered robot Figure 02 has joined BMW for work...

Although "Tiangong" is not in the same race as them, mainly focusing on solving common issues of embodied intelligence, it has also not escaped the fate of doing manual labor.

It is understood that it is initially conducting data collection and training in some scenarios, such as repetitive, tedious material handling and sorting, and conducting inspections and rescue operations in dangerous environments such as mines and construction sites...

The 2024 World Robot Conference will be held in the Beijing Economic and Technological Development Area on August 21, and those looking forward to the evolution of the "Tiangong" family can keep an eye out, as it is said that the Innovation Center will have a wonderful booth, where you can see the collective appearance of the "Tiangong" robot family, and there will also be a sub-forum.

One More Thing

Why do we let robots run on a treadmill? Let's guess (doge).

Answer: In addition to the higher demand for balance, the speed of the treadmill is controllable, which can test more accurately and objectively; the robot running experiment requires a larger venue, and the treadmill saves space.